Adam Daum
← All posts

Peter Steinberger Is Winning At AI

ailive-codingagentic-engineeringworkflow

Peter Steinberger Is Winning At AI

🤖 Heads Up! This post was written with AI assistance using Claude Code.

I've really been diggin Peter Steinberger lately. I wrote about a post about Steinberger's philosophy on building with AI, and some of my key takeaways. It Feels like he's winning. and I get it -- he sold his AI agent only social media platform, Moltbook to Meta. His open source project, OpenClaw, went viral, inspiring a movement. And he joined Open AI. I've been reading his blog (choc full of goodies). And I recently watched a live coding session he posted on YouTube as part of a blog post about adding a new feature,Arena, to another app he wrote with AI agents, sweetistics . He's embraced AI development tools, and open source, and the results are obvious. He's prolific. And, at heart, he's likable, he's genuine. He comes off as the type of builder who does it for the love of the game.

These are some of my takeaways from Peter's live coding session, and notes about how I've integrated some of his ideas into my workflow, especially talking out my prompts, building my projects by voice. this is the live coding session where he built and shipped a complete feature in roughly an hour. Blog posts are cleaned up. Live sessions are messy, honest, and full of the small decisions that actually define a workflow. Here's what stood out.

"No, I don't do TDD, fuck that. I write tests afterwards, it's the best way to get the model to correct itself." - Peter

TL;DR

  • 🛠️ CLIs over MCP. Small CLI wrappers keep the context window clean. MCP servers bloat it with noise.
  • 🧠 Don't write code — describe intent. Let the agent own the implementation and its own conventions.
  • ✅ Tests after, not before. Build the feature, then have the agent backfill tests against what actually exists.
  • 🎙️ Speak your prompts. Voice keeps you focused on intent instead of over-specifying implementation.
  • ⚡ Atomic commits to main. No feature branches. Small commits. Git is the safety net.
  • 📂 Stage files by name, never git add . Critical for multi-agent workflows where other processes may be writing to the repo.
  • 📋 Paste errors verbatim. Don't diagnose — the raw error is more useful than your interpretation of it.

MCP Is Unnecessary Complexity

I've never fully understood MCP. The models can read docs, figure out an API, and write the integration — the same way a human would. Why add a protocol layer?

Steinberger seems to agree with this. He uses small CLI wrappers instead of MCP integrations. They do the same job without bloating the context window with tool definitions and schemas. The one exception: Firecrawl for web search in Codex, since its built-in search isn't great. But I'm sure they'll get that worked out.

Maybe my perspective will change some day, but I don't need an abstraction layer between AI and my tools. You need the AI to use your tools the same way you would.

Let the Agent Own the Code

Steinberger doesn't write code in the session. He describes intent and reviews output. I'd been stuck trying to get AI to match my conventions — but why? If the agent understands its own code better than mine, let it write code the way it reasons about code. It'll maintain it better next time.

The developer's job is to describe intent clearly and evaluate whether the output works. Not to dictate style.

Tests After, Talk Instead of Type, Commit to Main

Three tactical choices that are easy to describe but hard to actually adopt:

Tests after. "I don't do TDD. I write tests afterwards." When an agent can generate both implementation and tests in seconds, writing tests first just locks you into an interface that's about to change.

Voice prompting. Steinberger speaks most of his prompts. Speaking forces you out of the weeds — you describe what you want the way you'd explain it to a colleague. I've started doing this and the difference is real.

Atomic commits to main. No branches, no PRs for solo work. Small commits directly to main. And never git add . — always stage specific files. In multi-agent work, a blanket add could stage another agent's in-progress changes.

What I've Changed

Since watching the session, I've adjusted my own workflow:

  1. 🎙️ Voice prompting by default (this one's a game changer. Feels weird at first, but damn, typing is slow .. haha)
  2. 🤝 Letting the agent own code style (shifting my output to something my agents can understand, not me.)
  3. 🛠️ CLI-first, no MCP (this wasn't new, but it's a validation. I spend less time investigating MCPs; that could change in the future, but )
  4. 📸 Pasting screenshots of errors, instead of explaining them.

Adam Daum is an agentic engineer and AI architect. He runs Weststack, LLC, an agentic AI and software engineering company, and writes about building practical AI solutions at adamdaum.com.